CI Loop: closed-loop analytics
The Composite Intelligence idea as a runnable demo: governed BI metrics, ML anomaly detection and forecasting, human-approved actions, and outcomes fed back as data.
The problem
Most analytics stacks stop at the dashboard. Someone notices a drop in sales, investigates, maybe acts, and nobody records whether the action worked. The organisation never learns.
What this project will show
A small but complete implementation of the CI loop on a synthetic retail sales dataset:
- Instrument: ingest orders, customers and products into bronze, silver and gold layers, with metric definitions as code and data quality checks.
- Model: daily forecasts and anomaly detection on key metrics, tracked in MLflow.
- Decide: when a metric breaks its expected range, generate a recommendation with the evidence attached.
- Act: a human approves or rejects it, and the decision is logged.
- Learn: outcomes are written back and compared against the forecast, closing the loop.
Milestones
- Synthetic data generator with realistic seasonality and injected anomalies
- Medallion pipeline with tests and data quality checks
- Forecasting and anomaly models with experiment tracking
- Recommendation and approval log
- Feedback dashboard: did the actions help?
Design principles
Local-first and reproducible: anyone should be able to clone the repository and run the whole loop with a single command.